fix(evals): 「暂不支持」只套功能类要求；伞状单元和条件性要求不挂问题；低密度投标文件；run_check 加 --v1 对比开关

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

evals/check/build.py

@@ -366,8 +366,18 @@ def build(tender: str, bid: str, style: str, run: str) -> None:
             if f.get("only_deviations"):
                 devs = [r for r in src if r["deviation"] != "无偏离"]
                 rows = [_row(k, r, columns) for k, r in enumerate(devs, 1)]
-                empty = {"no": "1", "requirement": "—", "answer": "—", "deviation": "无偏离"}  # 苏采云范本：没有偏离时注明「无偏离」
-                rows = rows or [{"cells": [empty.get(c, "") for c in columns], "anchor": columns.index("deviation") if "deviation" in columns else 0}]
+                empty = {
+                    "no": "1",
+                    "requirement": "—",
+                    "answer": "—",
+                    "deviation": "无偏离",
+                }  # 苏采云范本：没有偏离时注明「无偏离」
+                rows = rows or [
+                    {
+                        "cells": [empty.get(c, "") for c in columns],
+                        "anchor": columns.index("deviation") if "deviation" in columns else 0,
+                    }
+                ]
                 pool = [
                     _row(0, r, columns) for r in src if r["deviation"] == "无偏离"
                 ]  # 没列出的（默认接受）：注入时可以挑来写成偏离

evals/check/mutate.py

@@ -29,6 +29,11 @@ from build import rmb_upper  # noqa: E402
 
 OUT = EVALS / "out" / "check"
 VOCAB = {"dev": EVALS / "check" / "vocab-dev.json", "holdout": OUT / "_sealed" / "vocab-holdout.json"}
+# 「暂不支持」只对功能类要求说得通；证书、清单、服务义务、文档交付、改造范围这类不是功能（复核两批共拒掉 12 条）
+FUNCTIONAL = re.compile(r"支持|功能|具备|实现|可对|能够|可以|兼容|提供.{0,6}接口")
+NON_FUNCTIONAL = re.compile(r"证书|认证|清单|服务|培训|文档|资料|方案|范围|记录|报告|期限|质保|交付|人员|工期|成果")
+# 条件性要求：只对特定投标人适用（「如依法免税则须提供……」），规矩的投标人不适用，不能当作必须提供的材料
+CONDITIONAL = re.compile(r"(如|若|对于)[^，。；]{0,12}(免税|不需|无需|依法|属于|是)")
 DOC_REQ = re.compile(r"提供.{0,12}(方案|证明|报告|文件|承诺|截图|复印件|证书)")
 UNIT_DOWN = {"TB": ("GB", 1024), "GB": ("MB", 1024), "Gbps": ("Mbps", 1000), "s": ("ms", 1000), "秒": ("毫秒", 1000)}
 
@@ -38,6 +43,7 @@ class Mutator:
         self.spec, self.vocab, self.rng = spec, vocab, rng
         self.unit_defects: dict[str, dict[str, Any]] = {}
         self.rule_defects: list[dict[str, Any]] = []
+        self.umbrella: set[str] = set()
         self.vols = spec["volumes"]
 
     @property
@@ -140,8 +146,8 @@ class Mutator:
         if DOC_REQ.search(req_text):  # 「提供……方案 / 证明 / 报告」是文件要在（presence），不注入参数类问题
             return False
         numeric = tp.get("op") in (">=", "<=", "=") and tp.get("value")
-        if op == "D05" and numeric:  # 「暂不支持」只套在功能类要求上（尺寸、费用这类数值参数不会这么写）
-            return False
+        if op == "D05" and (numeric or not FUNCTIONAL.search(req_text) or NON_FUNCTIONAL.search(req_text)):
+            return False  # 「暂不支持」只套在功能类要求上（数值参数、证书、清单、服务义务不会这么写）
         if op == "D01":
             t["rows"].remove(r)
             self.drop_everywhere(units)
@@ -230,7 +236,19 @@ class Mutator:
                     row["cells"][dev] = "负偏离"
                     self.mark(row["responds"], "D06", "数值不满足并写负偏离")
                 else:
-                    param = tp.get("param") or (row["cells"][req][:12] if req is not None else "该功能")
+                    req_text = row["cells"][req] if req is not None else ""
+                    if not FUNCTIONAL.search(req_text) or NON_FUNCTIONAL.search(req_text):
+                        row["cells"][dev] = self.phrase("D06", "deviations")  # 非功能要求：写负偏离而不是「不支持」
+                        if note is not None:
+                            row["cells"][note] = ""
+                        self.mark(row["responds"], "D06", row["cells"][dev])
+                        if t["rows"] and "—" in t["rows"][0]["cells"]:
+                            t["rows"].clear()
+                        if self.col(t, "no") is not None:
+                            row["cells"][self.col(t, "no")] = str(len(t["rows"]) + 1)
+                        t["rows"].append(row)
+                        continue
+                    param = tp.get("param") or (req_text[:12] or "该功能")
                     row["cells"][ans] = self.phrase("D05", "phrases").format(param=param)
                     row["cells"][dev] = "负偏离"
                     self.mark(row["responds"], "D05", row["cells"][ans])
@@ -240,12 +258,26 @@ class Mutator:
                     row["cells"][self.col(t, "no")] = str(len(t["rows"]) + 1)
                 t["rows"].append(row)
 
-    def attachments(self, n: int, cats: dict[str, str]) -> None:
-        """删掉附件：资格条件 → D11，评分项 → D13，其他 → D02。"""
-        cands = [b for b in self.blocks if b.get("attach") and self.free(b.get("responds") or [])]
+    def attachments(self, n: int, cats: dict[str, str], texts: dict[str, str]) -> None:
+        """删掉附件：资格条件 → D11，评分项 → D13，其他 → D02。
+        伞状单元（「第二十二条规定的条件」这类总述，挂在 3 份以上附件上）不挂问题：查标报的是具体条目，和总述对不上；
+        条件性要求（「如依法免税则须提供」）规矩的投标人不适用，它的附件不删。"""
+        uses: dict[str, int] = {}
+        for b in self.blocks:
+            if b.get("attach"):
+                for u in b.get("responds") or []:
+                    uses[u] = uses.get(u, 0) + 1
+        self.umbrella = {u for u, k in uses.items() if k >= 3}
+        cands = [
+            b
+            for b in self.blocks
+            if b.get("attach")
+            and self.free([u for u in b.get("responds") or [] if u not in self.umbrella])
+            and not any(CONDITIONAL.search(texts.get(u, "")) for u in b.get("responds") or [])
+        ]
         self.rng.shuffle(cands)
         for b in cands[:n]:
-            units = b.get("responds") or []
+            units = [u for u in b.get("responds") or [] if u not in self.umbrella]
             kinds = {cats.get(u) for u in units}
             defect = "D11" if "qualification" in kinds else "D13" if "scoring" in kinds else "D02"
             self.mark(units, defect, "附件缺失", worded=False)
@@ -382,6 +414,10 @@ def main() -> None:
     tender = spec["tender"]
     units, _ = gold.units(tender)
     cats = {u.id: u.cat for u in units}
+    texts = {
+        u.id: "".join(str(x.get("t", "")) for x in u.raw.get("anchors") or []) + str(u.raw.get("note") or "")
+        for u in units
+    }
     optional = {u.id for u in units if u.optional}
     modes = unit_modes(spec)  # 注入前算：被删掉的行也要知道它本来该怎么响应
     m = Mutator(spec, json.loads(VOCAB[a.split].read_text()), random.Random(a.seed))  # noqa: S311 评测数据，不用于安全
@@ -390,13 +426,13 @@ def main() -> None:
             m.d01_block()
             m.row_ops(18)
             m.implicit_deviation(4)
-            m.attachments(4, cats)
+            m.attachments(4, cats, texts)
             m.compliance()
             for op in (m.p04, m.p05, m.p06, m.p07, m.p08, m.p09, m.p10, m.p12):
                 op()
         else:
             m.row_ops(4)
-            m.attachments(1, cats)
+            m.attachments(1, cats, texts)
             m.p04()
             m.p07()
     bid = a.clean if a.manifest_only else a.out
@@ -406,6 +442,10 @@ def main() -> None:
     for u in units:
         mode = modes.get(u.id) or ("na" if u.id in na else "unknown")
         entry: dict[str, Any] = {"cat": u.cat, "mode": mode, "optional": u.id in optional}
+        if u.id in m.umbrella and not a.manifest_only:
+            entry["expect"] = "unknown"  # 伞状单元：附件删了一部分，算不算问题看具体条目，评分不算
+            manifest_units[u.id] = entry
+            continue
         if u.id in m.unit_defects:
             entry.update({"expect": "defect", "reviewed": False, **m.unit_defects[u.id]})
         elif mode == "implicit":

evals/check/report.py

@@ -1,6 +1,6 @@
 """汇总一批投标文件的查标评分（evals/check/GUIDELINES.md §7）。先对每份跑过 score.py。
 
-    cd server && ../.venv/bin/python ../evals/check/report.py --run base <投标文件> [...]
+cd server && ../.venv/bin/python ../evals/check/report.py --run base <投标文件> [...]
 """
 
 from __future__ import annotations

evals/check/run_check.py

@@ -119,6 +119,7 @@ async def main() -> None:
     ap.add_argument("--checklist", default="fix3")
     ap.add_argument("--name", default="base")
     ap.add_argument("--baseline", action="store_true", help="只检索、不调模型的基线")
+    ap.add_argument("--v1", action="store_true", help="第 1 步的做法：提示词 v1、不带同一行的单元格（对比基线）")
     args = ap.parse_args()
     d = OUT / args.bid
     manifest = json.loads((d / "manifest.json").read_text())
@@ -128,6 +129,9 @@ async def main() -> None:
         pages = text_layer.extract(f"{args.bid}-{pdf.stem}", pdf.read_bytes())
         docs.append(match.BidDoc(uuid4(), n, pdf.stem, text_layer.body_sentences(pages)))
     ai_service.run_json = baseline_run_json if args.baseline else cached_run_json  # type: ignore[assignment]
+    if args.v1:
+        match.PROMPT = ("check.match", 1)
+        match.ROW_MATES = 0
 
     async def progress(_: float) -> None:
         return None